imwenyaot/imwenyaot-skills

Use when several finished runs have to become one table that survives review: ablation arms that differ in more than the thing under study, 消融臂不可比, 消融表, a baseline reproduced below its published number, an unfair comparison where one arm got a bigger tuning budget, how many seeds to run and what an error bar captures, 误差棒, seed 方差, and tracing a table row back to the commit and command that produced it. Do not use for diagnosing a single run that will not train, for drafting the prose around results, or for rendering a comparison as a figure.

Publication figure budgeting and production for academic papers, including 论文绘图/画图/架构图/结果图: reference-derived budgets, Elsevier/CVPR/ICCV/NeurIPS figures, diagrams, plots, result stitches, and publication-ready exports. Covers charts of every form — grouped bars, scatter, Pareto fronts, heatmaps — colour-blind-safe encoding, 600 dpi exports, and baseline-comparison panels. For an architecture figure it writes the structure as mermaid, derives the 生图提示词 an image model needs, checks each returned image against it, and corrects the draw.io or PowerPoint file traced from the accepted one. Do not use for language review, caption-only edits, or submission packaging.

Use when a paper has to get through a journal — its source, its upload packet, or its answer to the reviewers. Covers 期刊模板, 双栏 layout, 页数超限, and an elsarticle or IEEEtran source that will not compile; 投稿 and 投稿材料 for Elsevier Editorial Manager or the IEEE Author Portal, cover letter, highlights, required statements, EDICS, and the source zip; and 修回稿, 返稿, 审稿意见, 怎么回复审稿人 — a major or minor revision decision whose reviewer comments each need 逐条回应 in a point-by-point response letter. Do not use for writing or reviewing the paper's argument and prose, or for producing figure artifacts.

Staged setup and silent-failure diagnosis for neural network training. Use when you start a new training pipeline, when you review a training loop, or when a run goes wrong — the loss does not decrease, accuracy stalls, gradients explode or vanish, train and eval disagree, inference is wrong, or a one-batch overfit fails. Chinese triggers include 从零搭训练流程, 检查训练循环, loss 不下降, acc 卡住, 梯度异常, 训练验证不一致, and 推理错误. Do not use for experiment orchestration, ablation matrices, or result tables.

Academic-paper drafting, review, and polishing. Use when experiment results or an algorithm need a paper structure; to draft an introduction, abstract, related work, or contribution bullets; to polish manuscript LaTeX or a NeurIPS paragraph while preserving technical claims; to review logical self-consistency, experiment narrative, captions, figure-text linkage, and evidence-bound conclusions; to write the paragraph interpreting what a result means for the contribution; or to remove AI tone. Chinese triggers include 从零开始写论文, 写摘要, 写引言, 相关工作, 起草论文, 贡献列表, 润色论文 LaTeX, 去掉 AI 味, 不改变技术主张, 首句点题, 图表 caption, 证据自洽, and 终稿红线. Do not use for answering reviewer comments on a revision, generic editing, standalone figure production, or submission packaging.